Modal Strain Energy-Based Structural Damage Detection Using Convolutional Neural Networks
Round 1
Reviewer 1 Report
This is a much-improved paper from the last version. However, the lack of spatial plots in the results make it a little hard to find the key components of the paper. I would highly suggest adding back updated figures 5 and 6 from the last version.
All the best,
Austin.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
The authors present a structural damage detection method based on convolution neural network. The introductory material is sufficient and well presented. The method section must be improved by adding significant details like the exact distance in meters of the excitation point from the supports, the total weight of the structure, the weight of the accelerometers and cables attached to the structure, the sampling speed in samples per second of the acquisition equipment, and the bit depth of the acquired signals. Some flaws are present in the result section: the authors in Table 2 and 3 indicate that the natural frequency of the steel frame is around 3.906 Hz. However, in Figures 7 through 9, 5120 points were shown and the time label reads seconds as measurement unit. With such a sampling frequency (1 Hz), it is impossible to measure a natural frequency of 3.906 Hz since this would violate the Shannon sampling theorem. The authors should doublecheck the sampling frequency of the acquired data and correct either the time unit of Figures 7 through 9 or the frequencies in Table 2 and 3. If both these quantities are correct, data acquisition must be repeated with a sampling frequency of at least 10 Hz.
Some other minor errors are present in the text:
l.27 sudden > the sudden
l.38 of natural > in natural
l.43 In order to > To
l.51 formidable > not affordable
l.60 was proposed > is proposed
l.60 damage of a > damage in a
l.60 The numerical > Numerical
l.61 the experimental > experimental
l.67 witdth and > width, and
l.70 vibration signal > the vibration signals
l.96 Damage was > A damage was
l.97 were shown > are shown
l.106 width > a width
l.107 height > a height
l.122 2 convolution > two convolution
l.122 a pooling layers > one pooling layers
l.123 2 activation layer > two activation layers
l.123 and output > and one output
l.124 a softmax > one softmax
l.142 In order to > To
l.194 on real > on the real
l.220 multiple damage > multiple damages
l.223 was supplemented > were supplemented
l.230 has the ability of feature learning > can learn features
l.235 damage and > damage, and
Assuming that the time scale of the recorded data is fine, the conclusions are supported by the data and the discussion. As such, it is opinion of the reviewer that this work should be accepted after minor revision.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
The authors present a Convolutional Neural Networks (CNN) was used to extract the damage 10 features of a steel frame structure. As structural damage will induce changes of the modal 11 parameters of the structure, the convolution operation was used to extract the features of modal 12 parameters, and a classification algorithm was used to judge the damage state of the structure. - The paper has to be checked for English language grammar and spelling mistakes. Not all symbols are shown properly in the equations in my version. - The abstract should be revised highlighting the novelties. - The introduction has many references, but the new contributions are not clear. - Please double-check the symbols of all equations. - several relevant works should be added : 1- A computational approach for crack identification in plate structures using XFEM, XIGA, PSO and Jaya algorithm https://doi.org/10.1016/j.tafmec.2019.102240 2- Structural health monitoring using modal strain energy damage indicator coupled with teaching-learning-based optimization algorithm and isogoemetric analysis https://doi.org/10.1016/j.jsv.2019.02.017 3- Model Updating for Nam O Bridge Using Particle Swarm Optimization Algorithm and Genetic Algorithm https://doi.org/10.3390/s18124131 4- Damage Detection in Laminated Composite Plates Based on Local Frequency Change Ratio Indicator https://doi.org/10.1007/978-981-13-8331-1_71 - The authors should describe the number of sensors used in the experimental part also the strike position and number - The authors should describe the number of tests used into CNN and which data has good accuracy frequency or mode shape? - The conclusion should describe the advantage of the proposed applicationAuthor Response
Please see the attachment.
Author Response File: Author Response.pdf
This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.
Round 1
Reviewer 1 Report
@page { margin: 0.79in } p { margin-bottom: 0.1in; line-height: 115% }Structural Damage Detection with Modal Strain Energy using Convolutional Neural Networks
Overall, the paper is interesting but lacks any clear novelty in the field of SHM or machine learning. I have 2 major comments and several minor comments and feel this paper should be greatly improved to show the novelty of the paper to the field.
Major:
Line 58 states that “The application of CNNs to SHM provides a new intelligent method for structural damage detection. It is expected that CNNs can be applied to predict the locations and levels of damage in a structure.”, however, a simple search for “convolutional neural network structural health monitoring” in google scholar brings up several papers that do some form of this. I suggest the authors provide a more detailed review of what has been accomplished before in the field. Look at the references and papers that cite the recent paper below to get a good understanding of the state-of-the-art.
Real-time vibration-based structural damage detection using one-dimensional convolutional neural networks
Major 2: The contributions need to be clearly stated. As is, I see you are using a CNN to match a trained data set without experimental validation. This has been done thousands of times for hundreds of applications so without a cleat contribution it should be hard to convince the editor this should be published. This paper had excellent results, as would be expected from using one of the best pattern matching algorithms (CNN) to match data from an FEA model to itself.
Minor :
Figure 1 would benefit from showing a cross section of the beam. Also, this figure needs to be cleaned up, the fixity lines are not constant, the triangle is off and the numbers are not evenly centered over the elements.
The English is aqueduct, but the authors mess up tenses (present, past, future) a lot and I suggest they make a detailed review for these.
Line 115-118, remove only. It is hard to say its only damaged 90%. That defeats the use of the word only.
Datasets should not have a capital D in my view.
Are you collecting strain at each element (line 119)? If so, this seems a little obsessive. An additional study where you investigate the response for a limited number of inputs to the system would add novelty to the paper. A novelty that in my view this paper needs.
Is there data worth showing in figure 4? It seems to train in 100 epochs. Maybe put the y-axis in a log plot.
Reviewer 2 Report
The manuscript is poorly written and the results are not supported by any experimental data. Numerical simulation is not explained properly. No numerical signals/data are presented as claimed by the authors. The description/results presented in the manuscript looks too virtual.
The title of the manuscript is also incorrect. The very first sentence in the Abstract is also incorrect. I recommend the authors to take an extensive English correction & modify accordingly before submitting to any journal.
Finally, I do not recommend this manuscript for publication.
